Predicting the convergence of BiCG method from grayscale matrix images

抄録

The convergence of the BiConjugate Gradient (BiCG) method depends on its input matrices. We tried to predict the convergence of BiCG method by applying a Convolutional Neural Network to matrices that had been converted to grayscale images. Using 875 real non-symmetric matrices in the SuiteSparse Matrix Collection, we applied the 5-fold cross-validation method and were able to predict convergence with an average accuracy that exceeded 80\% for all cases in the test collection.